Table 3.

Alignment accuracy for global and local homologies of different evolutionary models implemented under the e2msa local alignment algorithm

	Alignment Accuracy
	[ AUC for F measure (%)]
Method	Global homology set			Local homology set
	parameterization			parameterization
	short	long	optimal	short	long	optimal
e2msa.afg	71.4	80.4	80.3	68.2	68.2	73.6
e2msa.aga	71.4	80.4	80.1	68.2	67.3	73.6
e2msa.aif	71.3	80.4	80.2	68.1	68.3	73.3
e2msa.tkf92	71.2	80.0	79.9	68.1	68.2	73.4
e2msa.afr	71.7	80.0	79.8	68.1	68.2	73.3
e2msa.aali	71.0	78.7	78.6	67.9	66.4	72.7
e2msa.tkf91	69.5	75.4	74.5	66.2	69.1	70.7
PHMMER (no filters) SSEARCH		78.7			72.9
(BLOSUM62, -11/-1)		80.0			71.7
NCBIBLAST		78.9			68.4
MSAProbs		81.7			NA
MUSCLE		80.8			NA

The “Global Homology set” is the one used in Fig. 7. The “Local Homology set” is the one used in Fig. 8. The e2msa algorithm was run in local mode, and with three different parameterizations: two at a fixed branch length (a short-branch and a long-branch parameterization, introduced in Fig. 7), and a variable optimal-time parameterization that uses for each homology the branch length that optimizes the probability of the sequences given the model. The rate parameters for all evolutionary model were obtained using the same training set “Pfam.seed.S1000.sto”. For all experiments, alignments are binned in 5 % identity groups, and the total F measure for one bin is calculated adding all alignments in that bin. In order to provide one single number, we report the area under the curve (AUC) for the F measure of alignments covering all identity ranges. For comparison, we provide results for other standard methods. Methods have been ranked by their combined performance in both sets. Methods such as MSAProbs and MUSCLE work only in “global” alignment mode, and they are not appropriate to detect local homologies

In bold, we indicate the best performing of the three alternative parameterizations